researcher edition

Researcher Edition

Adam Paszke, Sam Gross, Soumith Chintala, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Alban Desmaison, Andreas Kopf, Edward Yang, Zach Devito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy & Team

What is PyTorch?

What is PyTorch?

Ndarray library with GPU support

automatic differentiation engine

gradient based optimization package

Deep Learning

Reinforcement LearningNumpy-alternative

Utilities (data loading, etc.)

ndarray library•np.ndarray <-> torch.Tensor •200+ operations, similar to numpy •very fast acceleration on NVIDIA GPUs

ndarray library

Numpy PyTorch

ndarray / Tensor library

NumPy bridge

NumPy bridge

Zero memory-copy very efficient

NumPy bridge

Seamless GPU Tensors

automatic differentiation enginefor deep learning and reinforcement learning

PyTorch Autogradfrom torch.autograd import Variable

from torch.autograd import Variable

x = Variable(torch.randn(1, 10)) prev_h = Variable(torch.randn(1, 20)) W_h = Variable(torch.randn(20, 20)) W_x = Variable(torch.randn(20, 10))

PyTorch Autograd



i2h = torch.mm(W_x, x.t()) h2h = torch.mm(W_h, prev_h.t())

MMMM

PyTorch Autograd



i2h = torch.mm(W_x, x.t()) h2h = torch.mm(W_h, prev_h.t()) next_h = i2h + h2h

MMMM

PyTorch Autograd



i2h = torch.mm(W_x, x.t()) h2h = torch.mm(W_h, prev_h.t()) next_h = i2h + h2h

Add

MMMM

PyTorch Autograd



i2h = torch.mm(W_x, x.t()) h2h = torch.mm(W_h, prev_h.t()) next_h = i2h + h2h next_h = next_h.tanh()

Add

MMMM

Tanh

PyTorch Autograd




next_h.backward(torch.ones(1, 20))

Add

MMMM

Tanh

PyTorch Autograd

Neural Networks

Optimization packageSGD, Adagrad, RMSProp, LBFGS, etc.

Research workflows Pain Points Core Philosophy

(of PyTorch)Upcoming features

how does PyTorch deal with them?

what do they look like in

the deep learning space

Research Workflows

Research Workflow

idea / theory

Implement modelsTraining & Validation

Add results to paper

design experiments pick datasets / environments

Research Workflow

idea / theory




Literature Review

Research Workflow

idea / theory




Literature Review

I'm watching you

Work items in practice

Writing Dataset loaders

Building models Implementing Training loop

Checkpointing models

Interfacing with environments

Dealing with GPUsBuilding optimizers Building

Baselines

Work items in practice

Writing Dataset loaders

Building models Implementing Training loop

Checkpointing models


Dealing with GPUsBuilding optimizers Building

Baselines

Python + PyTorch - an environment to do all of this

Writing Data Loaders

• every dataset is slightly differently formatted


• every dataset is slightly differently formatted • have to be preprocessed and normalized differently

• every dataset is slightly differently formatted • have to be preprocessed and normalized differently • need a multithreaded Data loader to feed GPUs fast enough


PyTorch solution: • share data loaders across the community!


PyTorch solution: • use regular Python to write Datasets:

leverage existing Python code


PyTorch solution: • use regular Python to write Datasets:

leverage existing Python code Example: ParlAI


PyTorch solution: •Code in practice


PyTorch solution: •Code in practice

Core Philisophy Upcoming Features

Pain PointsResearch Workflows



Cars Video gamesInternet


Cars Video gamesPretty much every environment provides a Python API


Cars Video gamesNatively interact with the environment directly

Debugging• PyTorch is a Python extension

Debugging• PyTorch is a Python extension • Use your favorite Python debugger

Debugging• PyTorch is a Python extension • Use your favorite Python debugger • Use the most popular debugger:

Debugging• PyTorch is a Python extension • Use your favorite Python debugger • Use the most popular debugger:

print(foo)

Identifying bottlenecks• PyTorch is a Python extension • Use your favorite Python profiler

Identifying bottlenecks• PyTorch is a Python extension • Use your favorite Python profiler: Line_Profiler

Compilation Time• PyTorch is written for the impatient

Compilation Time• PyTorch is written for the impatient • Absolutely no compilation time when writing your scripts whatsoever

Compilation Time• PyTorch is written for the impatient • Absolutely no compilation time when writing your scripts whatsoever

• All core kernels are pre-compiled

Ecosystem• Use the entire Python ecosystem at your will

Ecosystem• Use the entire Python ecosystem at your will • Including SciPy, Scikit-Learn, etc.

Ecosystem• A shared model-zoo:

Linear style of programming• PyTorch is an imperative / eager computational toolkit


- Not unique to PyTorch


- Not unique to PyTorch - Chainer, Dynet, MXNet-Imperative, TensorFlow-imperative, TensorFlow-eager, etc.


- Not unique to PyTorch - Chainer, Dynet, MXNet-Imperative, TensorFlow-imperative, TensorFlow-eager, etc.

- Least overhead, designed with this in mind - 20 to 30 microseconds overhead per node creation - vs several milliseconds / seconds in other options

Go Through an example

The Philosophy of PyTorch

The Philosophy of PyTorch

• Stay out of the way • Cater to the impatient • Promote linear code-flow • Full interop with the Python ecosystem • Be as fast as anything else

Upcoming features

Distributed PyTorch• MPI style distributed communication • Broadcast Tensors to other nodes • Reduce Tensors among nodes - for example: sum gradients among all nodes

Higher order derivatives• grad(grad(grad(grad(grad(grad(torch.norm(x)))))))

Higher order derivatives• grad(grad(grad(grad(grad(grad(torch.norm(x))))))) • Useful to implement crazy ideas

Broadcasting and Advanced Indexing• Numpy-style broadcasting • Numpy-style indexing that covers advanced cases

JIT Compilation•Possible in imperative frameworks •The key idea is deferred or lazy evaluation

- y = x + 2 - z = y * y - # nothing is executed yet, but the graph is being constructed - print(z) # now the entire graph is executed: z = (x+2) * (x+2)

•We can do just in time compilation on the graph before execution

Computation Graphs

Deep Learning Frameworks

Imperative Frameworks

JIT Compilation

Lazy Evaluation

Computation Graphs



JIT Compilation




Add

MMMM

Tanh

Lazy Evaluation

Computation Graphs



JIT Compilation




Add

MMMM

TanhGraph built but not actually executed

Lazy Evaluation

Computation Graphs



JIT Compilation




print(next_h)

Add

MMMM

Tanh

Data accessed. Execute graph.

Lazy Evaluation

Computation Graphs



JIT Compilation

•A little bit of time between building and executing graph - Use it to compile the graph just-in-time

Add

MMMM

Tanh

JIT Compilation

Computation Graphs



JIT Compilation

•Fuse and optimize operations

Add

MMMM

Tanh

Fuse operations. Ex:

JIT Compilation

Computation Graphs



JIT Compilation

•Cache subgraphs

Add

MMMM

Tanh

I've seen this part of the graph before let me pull up the compiled version from cache

Compilation benefitsOut-of-order

execution

1 2 3 3 1 2

BatchNorm

ReLU

Conv2d

Kernel fusion Automatic work placement

Node 0 GPU 0

Node 1 GPU 1

Node 1 CPU

Node 0 GPU 1

Node 1 GPU 0

Node 0 GPU 0

http://pytorch.org Released Jan 18th 100,000+ downloads 950+ community repos 8200+ user posts 245 contributors

With ❤ from

http://pytorch.org

researcher edition

Documents